Explaining Realized Returns: Roll’s R2 40 Years Later -- by Bryan T. Kelly, Yinan Su, Chao Zhang

TEXT START: Can we explain asset price movements even with ex post observations of their drivers?

1. The Dissection

The paper reopens Roll’s R² problem and argues that conventional macroeconomic, cash-flow, expectations, and flow variables are weak explanations of realized equity returns. It then replaces coarse numerical proxies with AI-extracted reasons from Wall Street Journal market summaries, finding 30–40% out-of-sample explanatory power at daily and monthly frequencies.

Its real contribution is methodological: financial explanation is being shifted from human-designed factor models to machine-readable narrative data. The headline still overstates the achievement. Sixty to seventy percent remains unexplained, and “out-of-sample” is not necessarily real-time prediction when the explanatory text is produced after the market move.

2. The Core Fallacy

The decisive error under Discontinuity Thesis logic is confusing attribution with control. A narrative that explains why an asset moved does not establish a causal mechanism, create productive participation, or preserve the wage–consumption circuit.

Filtering out price language does not automatically remove post hoc rationalization, publication bias, shared information, or endogeneity. If the Wall Street Journal summary becomes available after the movement, the model is performing forensic reconstruction, not necessarily generating an actionable forecast. Statistical separation is not a time machine.

The paper therefore explains the market’s surface behavior more effectively without touching the system’s underlying death mechanism. Better explanation of financial claims does not rescue mass employment.

3. Hidden Assumptions

• Wall Street Journal summaries capture the relevant drivers rather than the drivers journalists selected, noticed, or were able to describe.
• Removing explicit price and return language removes leakage and narrative rationalization.
• Newspaper-stated reasons correspond to causes rather than socially accepted stories about causes.
• Out-of-sample performance implies practical ex ante availability and tradability.
• The identified themes remain stable across regimes and future information environments.
• AI-generated classifications are objective measurements rather than outputs shaped by prompts, filtering, labels, and model design.
• A higher R² represents deeper economic understanding rather than improved semantic compression.
• Greater legibility of returns has implications for broad economic welfare or human productive relevance.

The last assumption is the corpse beneath the laboratory table. Asset-market intelligibility and mass economic viability are different systems.

4. Social Function

This is a partial truth wrapped in prestige signaling and transition management.

The partial truth is real: standard factor models omit high-frequency information, and text can encode event-level conditions that aggregate macro variables miss. The prestige signal is the use of generative AI to convert elite financial prose into quantitative explanatory data. The transition function is quieter but more important: the paper demonstrates that interpretive financial labor can be standardized, extracted, and automated while presenting the result as a neutral improvement in R².

Under P1, the AI is not merely a research assistant. It is a machine for absorbing the narrative labor of journalists, analysts, and strategists. Ownership of the model, data, compute, and execution infrastructure accrues to Sovereigns. Human interpreters are pushed toward Servitor status or discarded.

5. The Verdict

Useful empirical work, but no defense of the post-WWII order. It shows that machine-processed narratives explain market returns better than conventional human-designed variables. That is a demonstration of cognitive automation and information concentration, not a restoration of productive participation.

The paper makes asset prices more legible while leaving the economic system’s terminal mechanism untouched. The market can become easier for machines to explain precisely as human labor becomes harder for the system to justify.